【作者】 李红霞；

【导师】 曾庆师；

【作者基本信息】 山东大学 ， 影像医学与核医学（专业学位）， 2024， 博士

【摘要】 研究背景随着数字化信息时代的来临,近些年各级医院基本完成了图像存储与传输系统(Picture archiving and communication systems,PACs)的安装与应用,存储了大量的医学影像图像,那么这些图像的存储怎么才能发挥更大和更重要的作用呢?人工智能(Artificial intelligence,AI)为此指明了一条清晰的道路。该系统具有更高的精度和计算能力,可以通过机器学习加强医学成像的分析,从而产生适合分层分类的多维数据。近些年来,随着AI的不断发展,机器学习,尤其是影像组学及深度学习在神经系统肿瘤方面的研究逐渐增多。垂体神经内分泌肿瘤(Pituitary neuroendocrine tumors,PitNETs),以往称垂体腺瘤(Pituitary adenoma,PAs),占颅内原发性肿瘤的15%左右,占到垂体原发性肿瘤的90%以上。根据肿瘤大小不同可以分为微腺瘤、大腺瘤、巨大腺瘤,其中微腺瘤是指肿瘤直径<10mm,大腺瘤是指肿瘤直径≥10mm,巨大腺瘤是指肿瘤直径≥40mm或肿瘤体积≥10cm3。按照内分泌激素是否异常,分为功能性和无功能性PitNETs。按照组织病理学细胞类型可以分为催乳素细胞瘤、促肾上腺皮质激素细胞瘤、生长激素细胞瘤、促卵泡激素-促黄体生成素细胞瘤、促甲状腺激素细胞瘤、零细胞瘤、双激素细胞瘤和多激素细胞瘤等。PitNETs多数为良性肿瘤,但仍有25%～55%的PitNETs具有侵袭性行为,这不仅可引起由于占位效应导致的头痛、呕吐、视野缺损等症状,还会导致手术切除困难从而影响预后。Ki67是一种细胞核蛋白,表达在全身多个系统中,因其与肿瘤增殖密切相关,所以有很多关于它的研究,并且在很多肿瘤的病理报告中会常规检测其表达水平。在2022年世界卫生组织(World Health Organization,WHO)关于神经内分泌肿瘤分类的第5次修订中,强调了形态学和免疫组织化学指标的严格相关性,其中就包括Ki67增殖指数的表达。以往研究表明,Ki67增殖指数与PitNETs的侵袭性和复发相关,Ki67表达水平≥3%被认为是肿瘤复发或进展的独立危险因素,Ki67增殖指数高,表明肿瘤切除术后需要更密切的随访或更早期的放疗。目前,结合影像组学和深度学习预测垂体瘤Ki67增殖水平的多中心研究较少。不同组织学亚型的PitNETs可引起不同的临床症状。临床上比较常见的有分泌催乳素、生长激素、促肾上腺皮质激素的肿瘤,可引起女性泌乳、男性性功能障碍、生长发育异常等,严重影响患者的生活质量。并且,不同的组织学亚型具有不同的治疗方法。多巴胺激动剂(Dopamine agonists,DA)药物治疗几乎是所有催乳素瘤(Prolactinoma,PRL)患者的一线治疗方法。以往研究表明,DA可以使90%的特发性高催乳素血症或微腺瘤患者和75%～80%的大腺瘤患者血清催乳素水平恢复正常,并且可以使90%以上的大腺瘤体积缩小。手术治疗只适用于一小部分对DA有耐药性的PRL患者,而对于其他组织学亚型的PitNETs,手术切除是一线治疗方法。因此,在制定治疗方案之前,区分组织学亚型是非常重要的。本研究的目的在于:(一)探讨基于多参数磁共振成像(Magnetic resonance imaging,MRI)的深度分割网络和影像组学术前预测PitNETs Ki67表达的可行性(二)探讨基于多参数MRI的深度分割网络和影像组学术前预测PitNETs组织学亚型的可行性第一部分:基于多参数MRI建立深度分割网络(cfVB-Net),并利用影像组学术前预测PitNETs Ki67表达研究目的1.基于多参数MRI建立PitNETs深度学习自动分割模型。2.利用影像组学特征术前预测PitNETs Ki67增殖指数的表达。3.结合临床因素、图像特征及影像组学评分构建Ki67临床预测模型。资料与方法本研究回顾性收集了自4个医学中心(医学中心1:559例,医学中心2:461例,医学中心3:129例,医学中心4:65例)的1214名PitNETs患者的术前MRI图像和术后病理资料。其中,来自3个医学中心(医学中心1、医学中心2、医学中心3)的1149例患者被随机分为训练集(Training set,n=804)和验证集(Validation set,n=345)。来自医学中心4的65例患者被分为外部测试集(External testing set,n=65)。在收集到的术前MRI图像中,随机选取155例患者的冠状位对比增强T1加权成像(Contrast enhanced T1-weighted image,CE T1WI)图像进行人工勾画,训练由粗到细VB-Net(the coarse to fine VB-Net,cfVB-Net)自动分割模型,并随机选取130例(医学中心1:80例,医学中心2:20例,医学中心3:20例,医学中心4:10例)患者,根据骰子相似系数(Dice similarity coefficient,DSC)、豪斯多夫距离(Hausdorff distance,HD)和平均对称表面距离(Average symmetric surface distance,ASSD)评估自动分割模型的分割性能。将所有患者分为Ki67高表达组(High Ki67 expression group,HG,Ki67 表达阳性率≥3%)和低表达组(Low Ki67 expression group,LG,Ki67表达阳性率<3%)。采用最小绝对收缩和选择算子(Least absolute shrinkage and selection operator,LASSO)方法筛选最优影像组学特征,并通过LASSO算法加权后所选特征的线性组合计算得出影像组学评分(Radscore)。然后采用Logistic回归、Bagging决策树(Bagging decision tree)和高斯过程(Gaussian process,GP)构建分类模型。采用受试者工作特征曲线下面积(Area under the receiver operating characteristic curve,AUC)评估模型在验证集和外部测试集中的性能。结合临床因素(年龄、性别)、影像学特征(肿瘤大小、Hardy’分级)和Radscore建立有效预测Ki67高表达发生概率的列线图模型。研究结果PitNETs深度分割模型对来自不同医学中心(DSC:0.870～0.920)、不同大小(DSC:0.723～0.930)和不同的Hardy’分级(DSC:0.775～0.930)的肿瘤均表现出良好的的分割效果。影像组学模型从多参数MRI图像中共提取1409个特征。LASSO共获得18个CE T1WI特征、15个T1WI特征和1 1个T2加权成像(T2-weighted image,T2WI)特征用于鉴别 Ki67 HG和LG由CE T1WI和T1WI构成的组合模型,在Bagging决策树中表现出最好的预测结果(AUC值:训练集0.927;验证集0.831;外部测试集0.825)。Logistic、Bagging决策树和GP模型在验证集中的AUC值分别为0.800、0.831和0.821。在外部测试集中,三种模型的AUC值分别为0.750、0.825和0.788。列线图中,年龄、Hardy’分级、CE T1 WI-Radscore、T1WI-Radscore是Ki67高表达的风险预测因子。决策曲线分析(Decision curve analysis,DCA)显示,列线图在概率阈值>0.10范围内获得Ki67高表达的最大净获益。结论基于多参数MRI的深度分割网络和影像组学分析在预测PitNETs Ki67表达方面具有良好的性能和临床应用价值。第二部分:基于多参数MRI的深度分割网络,利用影像组学特征术前预测PitNETs组织学亚型研究目的1.基于多参数MRI的深度分割网络,利用影像组学术前预测PitNETs多种组织学亚型。2.基于多参数MRI的深度分割网络,利用影像组学术前预测PitNETs催乳素瘤和非催乳素瘤。3.结合临床因素、图像特征及影像组学评分建立催乳素瘤临床预测模型。资料与方法本研究回顾性收集了 4家医学中心从2016年1月至2022年5月,共1206例PitNETs患者的医学影像图像。其中,来自医学中心1和中心2的患者按8:2的比例随机分为训练集和验证集,医学中心3和中心4的患者设置为外部测试集。所有患者均接受垂体T1WI矢状位及冠状位、T2WI冠状位、CE T1WI矢状位及冠状位检查。采用本研究第一部分中训练得到的cfVB-Net网络对PitNETs进行自动分割,并基于Python中的pyradiomics包从MRI中提取影像组学特征。通过LASSO回归选择最优特征,并利用通过LASSO计算得出所有特征的非零加权系数的线性组合计算每个患者的Radscore。为了预测组织学亚型,采用基于影像组学特征的高斯过程机器学习分类器,分别建立多分类(6类组织学亚型:催乳素细胞瘤-PRL、生长激素细胞瘤-GH、促肾上腺皮质激素细胞瘤-ACTH、促卵泡激素-促黄体生成素细胞瘤-FSH_LH、零细胞瘤-N、双激素或多激素细胞瘤-D_M)和二分类(PRL与non-PRL)GP模型。通过多因素Logistic回归分析,构建临床因素与Radscore相结合的临床-影像组学列线图。采用受试者工作特征曲线(Receiver operating characteristics curve,ROC)评价模型的性能。研究结果最终,PitNETs自动分割模型在4个医学中心的所有患者中获得的DSC平均值为0.888,从MRI图像中自动提取了 1409个影像组学特征。在肿瘤组织学亚型多分类(PRL、GH、ACTH、FSH_LH、N、D_M)模型中,通过特征筛选,最终在T2WI、T1WI和CET1WI三个序列中分别保留20、28、22个最优组学特征。T2WI序列的GP在训练集、验证集和外部测试集的AUC值最佳,分别为0.791、0.801和0.711。在PRL与non-PRL二分类模型中,通过特征筛选,最终在T2WI、T1WI和CET1WI三个序列中分别保留8、13、18个最优组学特征。T2WI和CET1WI构成的组合模型的GP表现最好,训练集、验证集和外部测试集的AUC值分别为0.936、0.882和0.791。在临床-影像组学列线图中,Radscore和Hardy’分级被确定为预测PRL表达的危险因子。结论基于多参数MRI的深度分割网络和影像组学特征在预测PitNETs组织学亚型方面表现出良好的性能和临床应用价值。更多还原

【Abstract】 BackgroundIn recent years,most hospitals,including different levels and sizes of hospitals have installed picture archiving and communication systems(PACs),with the coming of the digital information age.PACs can store a large number of medical images.So,how can we make these images play a more important role?Artificial intelligence(AI)shows a clear direction.With higher accuracy and powerful computing power,AI can enhance the analysis of medical image through machine learning,thus producing multi-dimensional data suitable for hierarchical classification.With the rapid development of AI,machine learning,especially radiomics and deep learning,has emerged in the study of nervous system tumors fastly.Pituitary neuroendocrine tumors(PitNETs;formerly known as pituitary adenoma,PAs)accounted for about 15%of intracranial primary tumors,and more than 90%of pituitary primary tumors.PitNETs are divided into microadenomas,macroadenomas,and giant adenomas according to their size.Microadenoma refers to the tumor with diameter<10mm,macroadenoma refers to the tumor with diameter≥10mm,and giant adenoma refers to the tumor with diameter≥40mm or volume≥10cm3.PitNETs are classified into functional and non-functional adenomas according to the pituitary hormone level.According to histopathological characteristics,PitNETs are divided into prolactinoma(PRL),adrenocorticoptrohic hormone tumor(ACTH),growth hormone tumor(GH),follicle-stimulating hormone_luteinizing hormone tumor(FSH_LH),thyroid stimulating hormone tumor(TSH),Null cell tumor(N),double hormone and multiple hormone cell tumor(D_M).Most PitNETs are benign tumors,but 25%～55%of PitNETs have aggressive behavior.These tumors not only cause headache,vomiting,visual field defect and other symptoms caused by the occupying effect,but also lead to surgical resection difficulties,which affect the prognosis.Ki67 is a nuclear protein that is expressed in multiple systems throughout the body.Because Ki67 is associated with tumor proliferation,it has been studied extensively and its expression levels are routinely measured in pathological reports of many tumors.The 5th revision of the World Health Organization(WHO)Classification of neuroendocrine tumors in 2022 highlighted the strict correlation between morphological and immunohistochemical indicators,including the expression of the Ki67 proliferation index.Previous studies have shown that Ki67 proliferation index is associated with more aggressive tumor behavior and recurrence of PitNETs.Ki67 expression level≥3%is considered an independent risk factor for tumor recurrence or progression.High Ki67 proliferation index indicates the need for closer follow-up or earlier radiotherapy after tumor resection.At present,there are few multi-center studies predicting the Ki67 expression of PitNETs combining radiomics and deep learning.Different histological subtypes of PitNETs can cause different clinical symptoms.Clinically,more common tumors secreting PRL,GH,ACTH can cause female lactation,male sexual dysfunction,physical development abnormalities,seriously affecting the life quality of patients.Moreover,different histological subtypes have different treatment methods.Dopamine agonists(DA)medication is the first-line treatment for almost all patients with PRL.Previous studies have shown that DA can normalize serum prolactin levels in 90%of patients with idiopathic hyperprolactinemia or microadenoma and 75%～80%of patients with macroadenoma,and can shrink more than 90%of macroadenoma volume.Surgical treatment is only appropriate for a small percentage of DA resistant PRL patients.For other histological subtypes of PitNETs,surgical resection is the first-line treatment.Therefore,it is very important to distinguish histological subtypes before developing the treatment plan.The purpose of this study:(Ⅰ)To investigate the feasibility of preoperatively predicting Ki67 expression in PitNETs by deep segmentation network and radiomics based on multi-parameter MRI.(Ⅱ)To investigate the feasibility of preoperatively predicting histological subtypes of PitNETs by deep segmentation network and radiomics based on multi-parameter MRI.Part I:Preoperatively Predicting the Ki67 Expression in PitNETs Using cfVB-Net and Radiomics Based on Multi-parameter MRI1.To establish a deep learning automatic segmentation model of PitNETs based on multi-parametric MRI.2.Preoperatively predicting the Ki67 expression in PitNETs using radiomics.3.The Ki67 prediction model was constructed by combining clinical factors,image features and radscores.Materials and MethodsThis retrospective analysis included 1214 PitNETs from 4 medical centers(centerl:n=559,center2:n=461,center3:n=129,center4:n=65)with complete preoperative MRI and postoperative pathology.1149 patients from medical center1-3 were divided into the training(n=804)and validation set(n=345)randomly,and 65 patients from medical center4 were assigned to the external testing set.155 cases were randomly selected to manually delineate the lesions to train the coarse to fine VB-Net(cfVB-Net)auto-segmentation modal,and 130 cases were selected to evaluate the performance of the segmentation model in accordance with the dice similarity coefficient(DSC),the hausdorff distance(HD),and the average symmetric surface distance(ASSD).All patients were divided into high Ki67 expression group(HG,Ki67 expression≥3%)and low expression group(LG,<3%).The least absolute shrinkage and selection operator(LASSO)regression method was applied for selecting the optimal radiomics features,and Radscore was calculated by linear combination of selected features weighted by LASSO algorithm.Then,logistic,bagging decision tree and gaussian process were performed to build the classification model.The performance of the model was evaluated with the areas under the receiver operating characteristic curve(AUCs)in the validation and external testing sets.Clinical factors(age,gender),imaging features(tumor size,Hardy’ grade)and radiomics score(Radscore)were combined to establish a nomogram that effectively predicted the Ki67 expression.ResultsThe cfVB-Net segmentation model showed good performance for different medical center(DSC:0.870-0.920),tumor sizes(DSC:0.723-0.930)and Hardy’ grades(DSC:0.775-0.930).A total of 1409 features were extracted from PitNETs MRI.Through LASSO,18 optimal features in the contrast enhanced(CE)T1WI,15 features in the T1WI,and 11 features in the T2WI were obtained for differentiating between Ki67 HG and LG.The best results were presented in the bagging decision tree when combined CE T1WI and T1WI(AUCs:training set,0.927;validation set,0.831;and external testing set,0.825).The AUCs of the logistic,bagging decision tree and gaussian process model were 0.800,0.831 and 0.821 in the validation set,respectively.In the external testing set,the AUCs were 0.750,0.825 and 0.788,respectively.In the nomogram,age,Hardy’ grade,CE T1WI-Radscore,and T1WI-Radscore were risk predictors of Ki67 high expression.The decision curve analysis showed that the nomogram achieved the greatest net benefit in the range of threshold probability>0.10 for patients at high risk of Ki67 expression.ConclusionDeep segmentation network and radiomics analysis based on multi-parameter MRI exhibited good performance and clinical application value in predicting the Ki67 expression in PitNETs.Part Ⅱ:Preoperative Identification of Histological Subtype in PitNETs Using cfVB-Net and Radiomics Based on Multi-parameter MRI1.To identificate multiple histological subtypes of PitNETs using radiomics and cfVB-Net based on multi-parameter MRI.2.To identificate PRL and non-PRL in PitNETs using radiomics.3.A nomogram model for predicting PRL was established by combining clinical factors,image parameters and Radiomic features.Materials and MethodsA total of 1,206 patients with PitNETs from January 2016 to May 2022 were retrospectively enrolled.The patients from centerl and center2 were randomly divided into training and validation sets according to the ratio of 8:2,and patients from center3 and center4 were designed as external testing set.All subjects accepted the pituitary TIWI,T2WI and contrast-enhanced(CE)T1WI examination.A cfVB-Net network was used to automatically segment PitNETs,and radiomics features were extracted from the MRI based on the pyradiomics package in Python.The optimal features were selected through the least absolute shrinkage and selection operator(LASSO)regression algorithm,and the radiomics score(Radscore)of each patient was calculated as the linear combination of each feature weighted by the non-zero coefficient from LASSO.To predict histological subtypes,Gaussian process(GP)machine learning classifier based on radiomics features were performed.Multi-classification(six-class histological subtype)and binary classification(PRL vs.non-PRL)GP model was constructed.Then,a clinical-radiomics nomogram combining clinical factors and Radscores was constructed by multivariate logistic regression analysis.The performance of the models was evaluated using receiver operating characteristic(ROC)curves.ResultsThe PitNETs auto-segmentation eventually achieved a Dice similarity coefficient of 0.888 in all patients from 4 centers.A total of 1,409 radiomics features were automatically extracted from the MRI.In Multi-classification model,20,28 and 22 features were selected as the optimal features in T2WI,T1WI and CE T1WI,respectively.The GP of T2WI got the best area under the ROC curve(AUC),with 0.791,0.801 and 0.711 in the training set,validation set and external testing set,respectively.In binary classification model,8,13 and 18 features were selected as the optimal features in T2WI,T1WI and CE T1WI,respectively.The GP of T2WI combined with CE T1WI demonstrated good performance,with the AUC of 0.936,0.882 and 0.791 in the training set,validation set and external testing set,respectively.In the clinical-radiomics nomogram,Radscores and Hardy’ grade were identified as predictors of PRL expression.ConclusionMachine learning and Radiomics analysis based on multiparameter MRI exhibited good performance and clinical application value in predicting the PitNETs histological subtypes.更多还原